Abstract: Dataset distillation compresses large datasets into smaller synthetic coresets that retain performance with the aim of reducing storage and computational burdens of processing an original, entire dataset. The present disclosure provides an improved algorithm that uses a non-deterministic feature approximation of neural network Gaussian process (NNGP) kernels, or other trained kernels, that reduces a kernel matrix computation to O(|S|). When combined with a modified Platt scaling loss, the disclosed algorithm can provide at least a 100-fold speedup over a Kernel-Inducing Points (KIP) algorithm and can run on a single graphics processing unit. The disclosed Random Feature Approximation Distillation (RFAD) algorithm can perform competitively with other dataset condensation algorithms in accuracy over a range of large-scale datasets, both in kernel regression and finite-width network training. The disclosed techniques can be effective on tasks such as model interpretability and data privacy preservation.

Abstract: System, methods, and other embodiments described herein relate to training a scene simulator for rendering 2D scenes using data from real and simulated agents. In one embodiment, a method includes acquiring trajectories and three-dimensional (3D) views for multiple agents from observations of real vehicles. The method also includes generating a 3D scene having the multiple agents using the 3D views and information from simulated agents. The method also includes training a scene simulator to render scene projections using the 3D scene. The method also includes outputting a 2D scene having simulated observations for a driving scene using the scene simulator.

Abstract: A controller for an autonomous vehicle is trained using simulated paths on a roadway and simulated observations that are formed by transforming images previously acquired on similar paths on that roadway. Essentially an unlimited number of paths may be simulated, enabling optimization approaches including reinforcement learning to be applied to optimize the controller.

Abstract: Systems and methods described herein relate to vehicular navigation and localization. One embodiment extracts perceptual features from sensor data; extracts unrouted-map features from unrouted map data; combines the perceptual features and the unrouted-map features to produce first combined features data; outputs, based at least in part on the first combined features data, parameters of a probability distribution for one or more steering trajectories that are available to a vehicle; and performs a localization of the vehicle based, at least in part, on the parameters of the probability distribution.

Abstract: Systems and methods described herein relate to vehicular navigation and localization. One embodiment extracts perceptual features from sensor data; extracts unrouted-map features from unrouted map data; combines the perceptual features and the unrouted-map features to produce first combined features data; outputs, based at least in part on the first combined features data, parameters of a probability distribution for one or more steering trajectories that are available to a vehicle; and performs a localization of the vehicle based, at least in part, on the parameters of the probability distribution.